How we use AI
Ripple Research is an AI-for-Good narrative analytics organisation. We have applied AI to public-interest research since 2020, and were among the first organisations to bring it to the study of public discourse on health, climate and democracy.
We have built our own AI tools for narrative research since 2020, before large language models were in general use
Ripple Research began in 2020. In the middle of the pandemic, social data had become the richest record of what people believed and the hardest to read, and we built tools to pull signal from that unstructured mass. We have worked as a firm since 2021, and our team co-authored the peer-reviewed study of the World Health Organization's Early AI-supported Response with Social Listening (EARS) platform, which applied machine learning to the COVID-19 infodemic.
Those first tools have grown into a proprietary research environment, built to serve our mission and configured for each project's question and languages. Each project draws on the same methods in a different combination.
Collection
Run proprietary search lexicons across defined, publicly available sources and archive every source.
Set the scope, and design and test the lexicons.
Structuring
Identify speakers and the people they mention, and flag statements that echo known narratives.
Read, parse and code the material.
Analysis
Measure how volume and networks of influence change over time.
Decide which questions to put to the data, and test whether the answers stand up.
Review
Cross-check coding and draft findings, using our LLM council when a project calls for it.
Resolve every disagreement and make the judgement calls.
Synthesis
Keep every finding traceable to its archived source.
Organise the evidence into frameworks and concepts, and write the findings and recommendations.
Our team turns the evidence into frameworks and strategy
Ripple Research brings together analysts, strategists, and policy and communications experts. They develop the concepts and frameworks that shape how we read public discourse, including Meet-and-Move, Beef to Debate, anxiety capture and narrative cartography, and turn what the research finds into strategy. Native speakers read every source in its original language.
Our LLM council adds independent review when a project calls for it
The LLM council is our approach to reviewing coded material and draft findings with a panel of large language models from different developers. Each model reviews the same material independently, and our team examines where they disagree. We bring the council in as each project requires.
Research shows that a single model acting as judge carries biases, including a preference for its own outputs (Zheng et al., 2023; Panickssery et al., 2024), and that panels and debates across several models give more accurate and less biased judgements (Du et al., 2024; Verga et al., 2024).
We work with leading universities to keep our methods rigorous and to share them
We work with universities and academics, including the Fletcher School at Tufts University, Texas A&M University and Maastricht University. These partnerships bring academic rigour to our work, and we share our methods with their research teams.
We protect data by working from public sources and keeping confidential material out of AI systems
The enterprise AI services we use do not train on the data we give them. Personal data is limited to public statements by people acting in a public or professional role. We comply with UK GDPR and EU GDPR, agree how AI will be used before each project begins, and provide an AI impact assessment on request.
Questions about how we use AI?
We are glad to take you through our approach and the safeguards that apply to any engagement.